Saturday, 5 September 2026
Artificial intelligence has generated no shortage of ambitious insurance projects, but the harder question for carriers is increasingly straightforward: where is it producing measurable business value in day-to-day operations? CURE Insurance provided a practical answer at AI4 2026 in Las Vegas, outlining how it has moved selected AI applications from experimentation into live insurance processes covering documents, claims, customer service and language translation. The company’s experience suggests that successful insurance AI is less about deploying the most sophisticated model available and more about selecting narrowly defined problems, integrating the technology into existing operations, measuring the result and ensuring employees trust the system enough to use it.
Douglas Benalan, Chief Information Officer and head of digital transformation at CURE Insurance, described four elements underpinning the company’s approach: identifying business value, building the appropriate solution, measuring whether it produces the expected result and establishing sufficient trust among employees, customers and management. That framework has shaped a technology programme that deliberately avoids treating AI as the answer to every operational problem. In a highly regulated industry built around complex workflows, sensitive data and legacy technology, CURE’s approach has been to begin with processes that contain substantial manual work but comparatively manageable risk.
The company’s starting point has been its frontline workforce. Rather than identifying applications exclusively within the technology department, CURE worked with operational employees and their supervisors to map repetitive tasks, areas susceptible to manual mistakes, processes involving significant delays and workflows that sit outside core insurance systems. Potential projects were then assessed against factors including technical capability, integration, governance, auditability and whether the company should develop the solution internally or purchase it externally.
That process is important because insurance companies frequently operate with multiple generations of technology. A technically impressive AI system can still fail if employees must leave their established workflow to use it. CURE therefore places considerable emphasis on integration with core systems so that AI becomes part of the employee’s normal working environment rather than another separate application competing for attention.
One of CURE’s earliest production applications involved insurance documents required from customers seeking particular premium reductions. Employees previously had to examine unstructured healthcare-related documents, which could contain multiple pages and appear in many formats, before manually comparing several pieces of information with the insurer’s systems. CURE introduced automated document processing to identify and extract the relevant information. According to Benalan’s presentation, the system initially processed approximately 55% of cases successfully when first introduced several years ago. Continued training and refinement subsequently increased that level to around 75%, with the company now reporting performance of approximately 90% for the process.
Instead of customers waiting for employees to review documents manually, qualifying information can increasingly be processed much closer to real time. The example demonstrates an important characteristic of enterprise AI deployment: production systems do not necessarily begin with exceptional performance. They can improve through repeated exposure to real cases, user feedback and better handling of unusual documents. The business case therefore depends partly on whether an organisation can create a controlled environment where the system is allowed to improve without introducing unacceptable risk.
CURE has also concluded that some problems should not use AI at all. Straightforward repetitive tasks can often be handled more cheaply and predictably through conventional software or robotic process automation. Generative AI introduces additional computing expense as well as governance requirements, meaning that using it for a problem already well served by deterministic software can unnecessarily increase both cost and complexity. The distinction is particularly relevant as companies face growing bills associated with model usage. An insurer deploying AI across millions of transactions needs to understand not only whether a model can complete a task but whether doing so represents the most economical approach.
CURE therefore reserves more advanced AI for areas where interpretation, summarisation or decision support genuinely adds value. Another design principle is the ability of an AI system to recognise uncertainty. In insurance, a model that produces a confident answer despite insufficient information can be considerably more dangerous than one that declines to proceed. CURE’s approach therefore includes situations in which the system should identify that confidence is inadequate and return the case to a human employee.
That human fallback is central to its production strategy. The company does not regard an AI process as sufficiently resilient if normal operations collapse whenever the AI service becomes unavailable. Employees need to understand what the technology has completed, what remains outstanding and how to continue processing the case manually if required. Business continuity therefore becomes part of AI architecture rather than an issue considered after deployment.
Claims operations provide several examples of where CURE is extending this approach. One involves extracting structured information from unstructured material such as handwritten notes and poorly scanned documents. Instead of claims employees spending time locating individual details inside difficult source material, AI can prepare the relevant information for review. The insurer is also using AI to summarise extensive medical records. Claims files can contain hundreds of pages of healthcare information, requiring employees to identify the relatively small proportion that is relevant to the insurance decision. AI agents can help extract and organise the information around predefined attributes, reducing the amount of manual reading while leaving the eventual judgement with employees.
CURE has attempted to reduce implementation risk by giving employees access to controlled testing environments before applications move into production. Users can experiment with different prompts, unusual documents and edge cases to understand both what the model does well and where it fails. This also helps employees gain confidence before the technology becomes embedded in live operations. The company additionally uses adversarial testing in which cross-functional teams deliberately challenge applications with difficult, unexpected or incorrect inputs. The purpose is not to demonstrate that the technology performs perfectly, but to discover weaknesses before policyholders or employees encounter them in live systems.
That testing philosophy is particularly relevant for generative AI, where outputs can vary and models can produce apparently convincing but inaccurate information. Insurance carriers therefore need governance capable of identifying not only obvious technical failures but subtler cases where an AI-generated answer appears plausible despite being wrong.
Another area moving into production is the first notification of an insurance claim. When a customer reports an accident, claims employees typically gather a large amount of information during the initial telephone conversation and enter it into internal forms. That administrative process can extend the duration of the call and divide the employee’s attention between the customer and data entry. CURE is using AI to analyse the conversation as it takes place and populate relevant fields automatically. The employee can then review the extracted information before accepting it, preserving human oversight while reducing the amount of manual typing required during the call.
The objective is to allow claims staff to concentrate more heavily on the customer at a moment when the policyholder may already be dealing with the stress of an accident. This represents a wider opportunity across insurance contact centres. Rather than replacing customer-service employees, AI can increasingly operate in the background by transcribing conversations, extracting information, preparing documentation and prompting employees when something requires attention. Human staff retain responsibility for the interaction while the technology handles more of the underlying administration.
Language translation provides one of CURE’s clearest examples of measurable financial returns. The insurer has introduced real-time translation to allow an employee to communicate with policyholders speaking languages the employee does not understand. Spoken communication can be translated during the interaction rather than requiring a separate human interpreter for every conversation. CURE initially focused on Spanish because of customer demand before expanding the capability to additional languages.
Benalan reported that the translation project has generated roughly three times its cost in returns while cutting translation expenditure by approximately 60% to 70%. The result provides an example of an AI deployment whose financial performance can be measured directly rather than inferred from general improvements in productivity. The value extends beyond cost reduction. Immediate translation can shorten customer-service interactions, reduce delays associated with obtaining interpreters and allow a wider group of employees to serve policyholders who speak different languages. For an insurer operating across diverse markets, that potentially improves both staffing flexibility and customer accessibility.
CURE’s experience also highlights why measuring AI programmes has become increasingly important. Enterprise organisations can accumulate large numbers of pilots without knowing which ones materially improve the business. Benalan argued that an initiative should not continue receiving substantial investment unless management can establish an appropriate performance measure. That does not mean every project must be evaluated entirely through direct cost savings. Depending on the application, relevant measures might include processing time, error rates, employee capacity, claims recovery, customer satisfaction or the number of cases completed without manual intervention. The important point is that the expected outcome needs to be defined before the technology is scaled.
Subrogation is another area in which CURE has been increasing its use of AI. The process involves identifying situations where another party may ultimately be responsible for costs that the insurer has initially paid. Traditionally, significant manual work can be required to examine claims files, assess recovery potential and prepare the documentation necessary to pursue reimbursement. CURE has been expanding technology that can analyse claims information and identify cases with stronger recovery potential, allowing specialists to concentrate on the claims most likely to justify further action.
The significance is that this application can influence the insurer’s financial result directly. Faster administrative processing is useful, but recovering additional money that would otherwise have been missed has a clearer connection to profitability. It illustrates the move from AI as a productivity tool towards AI being embedded in economically consequential insurance processes.
However, CURE’s presentation repeatedly returned to the importance of employee adoption. The company argues that even technically successful AI will produce little benefit if claims staff, customer-service teams or other operational employees do not trust it enough to incorporate it into their work. Employees therefore need to understand how the system reached its result, when they should question it and how it affects their own responsibilities. CURE has attempted to involve frontline staff during development rather than presenting them with completed systems created elsewhere in the organisation.
That participation can also expose operational details that technology teams might otherwise overlook. Psychological security forms another element of the implementation strategy. Employees need confidence that identifying errors, challenging outputs or suggesting improvements will not be treated as resistance to transformation. The organisation benefits when users actively search for weaknesses because those employees often understand the underlying process better than the teams building the technology.
This illustrates why enterprise AI is increasingly becoming a management issue rather than merely an information technology project. Legal teams, operational specialists, compliance staff, engineers and business leaders need to participate together because each sees different forms of risk. Customer trust creates an additional requirement. Policyholders need confidence that sensitive information remains protected and that automated systems do not introduce unfair or unexplained treatment. Management, meanwhile, needs sufficient auditability and monitoring to demonstrate that systems remain within the company’s risk tolerance.
Insurance is particularly sensitive to these issues because AI can potentially influence decisions involving premiums, claims and access to coverage. Governance therefore needs to continue after deployment. A system that performed correctly when launched can behave differently as customer behaviour, source data or underlying models change.
CURE’s experience points towards a more pragmatic stage in the insurance industry’s adoption of artificial intelligence. The debate is becoming less about whether insurers should experiment with AI and more about identifying the individual processes where the economics, operational benefits and risks justify deployment. The most successful applications may initially look less dramatic than fully autonomous underwriting or claims settlement. Extracting information from documents, summarising medical files, assisting claims employees during telephone calls and translating customer conversations are comparatively narrow applications, yet these are precisely the types of repetitive, high-volume activities where measurable improvements can accumulate across an insurance operation.
They also provide organisations with something strategically important: experience operating AI in production. Every successful deployment creates knowledge about governance, system integration, employee behaviour, model limitations and cost. That institutional capability can later support more consequential applications. For insurers, this may prove more valuable than attempting to leap immediately towards autonomous AI. A controlled system that solves a defined problem, integrates with existing operations and produces measurable returns can provide a stronger foundation than an ambitious project that never moves beyond demonstration.
CURE’s case therefore reinforces a broader lesson emerging from AI4 2026. Enterprise AI is entering a stage where novelty matters less than execution. The companies generating tangible value are increasingly those willing to define a narrow business problem, involve the people who actually perform the work, choose the simplest suitable technology, test it aggressively and measure what happens after deployment.
In insurance, where trust and continuity are fundamental, that discipline may ultimately determine which AI programmes progress from experimentation to core infrastructure. The technology itself is advancing rapidly, but measurable return depends on something considerably less fashionable: selecting the right problem and making the system work reliably in the real business.
Source: CIJ.World Research & Analysis Team
